基于修复素材优选的图像修复研究

With the rapid development of Internet technology, Chinese digital media industry enters the high-speed development period, a variety of media application technologies (film and television production, satellite map search, online images display) have entered into life. As a key technology of them, image completion technology offers important technical support to media application and has a very broad application prospect. However, the key problems are not solved effectively. “the accuracy of material image retrieval is low”, “it’s difficult to extract the material regions exactly” and “there are obvious artificial details in the process of completion” are existing problems of traditional methods. For the existing problems of traditional methods, our research focuses on searching problems of material image optimized retrieval, exact material region extraction, high-quality image completion. We propose joint entropy minimization based material image optimized retrieval model, multi-scale detail enhancement and multi-region clustering based exact material region extraction model, gradient field optimization of completion boundary based image completion model. Based on above models, we set up an application platform of completion material optimized retrieval based image completion, verify theoretical algorithm, to promote the model into practice and applied to digital media industry.

随着互联网技术的蓬勃发展，我国的数字媒体产业进入高速发展期，多种媒体应用技术(影视制作、卫星地图搜索、网络图片展示等)已经进入到日常生活中。图像修复技术作为其中的一项关键技术为媒体应用提供了重要的技术支持，并有着非常广泛的应用前景。然而，修复技术中的关键问题并没有得到有效地解决，素材图像检索正确率低、无法精确地提取素材区域、修复过程人为痕迹明显都是传统方法中存在的问题。本申请项目针对传统技术中存在的问题，重点研究素材图像优选、素材区域精确提取、图像高质量修复等相关问题。提出了基于图像联合熵最小化的素材图像优选方法、基于多尺度区细节加强和多区域聚类的素材区域精确提取方法、基于修复边缘梯度场最优化的图像修复方法，并在此基础上实现一个基于修复素材优选的图像修复应用平台，验证理论算法，促进结果走向实用，应用于数字媒体产业。

图像修复技术可以直接为文化、生活、商业等领域提供直接的技术支持，例如影视制作、网络图片展示、卫星地图搜索等。尤其，多选择性修复利用更加丰富的语义内容对缺失区域进行填充，体现了非常好的实际应用价值。然而，修复结果语义不匹配、边缘修复痕迹明显等问题显著地降低了修复的质量。针对上述问题，本项目提出了多特征联合的素材图像优选模型，提出了优化式多区域聚类的素材区域精确提取模型，提出了修复边缘梯度场最优化的图像修复模型。同时，利用多种理论模型互相验证的方法，确保了理论模型的正确性。在此基础上，搭建了基于修复素材优选的图像修复应用平台，并基于数十万张图像的图像库进行了充分实验，确保了平台在数字媒体领域的有效性。除此之外，模型和平台也在生命科学领域、教育文化领域进行了初步应用。较为理想的应用结果证明了模型和平台的有效性，也充分说明模型和平台具有广阔和潜在的应用价值。